Research On I-vector Based Speaker Normalization For Speech Recognition

Speech recognition speaker normalization subtractive primary purpose ofrandom differences between speakers, improving the constant parameters, filteringpersonal characteristics in the process to obtain valid information with linguisticmeaning. Another effect is reflected in the different pronunciation of the timerecording mode (formal, differences and tensions, etc.) to eliminate differences.i-vector speaker recognition is the more effective method is more popular inrecent years, modeling idea. It can better reflect the personality differences betweenthe speaker, an important advantage of this remarkable feature, both for speakerrecognition or validation of talking people showed its effectiveness. We can use thesedifferences in speech recognition and clustering. After clustering, according to thisclustering information for speaker normalization should be able to obtain betterspeech recognition result.Based on the above ideas, this article will i-vector used in the acousticcharacteristics of speech recognition speaker normalization: First training speech dataextraction feature vector i-vector and use unsupervised clustering algorithm LBG,LBG algorithm without two types of supervised clustering reflects the gendercharacteristics of men and women. Then the maximum likelihood training of all kinds,respectively, using a linear transformation to achieve speaker adaptation trainingspeaker normalization. The characteristics of the transformed speech for speakerrecognition training and recognition, the experimental results show that the methodcan improve the performance of speech recognition.